Track updated status for each vector in multivector.

[pelesh]

Vector class object is a multivector that allows access to individual vectors separately. It is plausible to expect that individual vectors could be updated individually and some of them may have device, while others may have host memory space updated. However, current implementation of the Vector class allows only for setting update flags for all vectors together, regardless of how they are accessed and updated. This is a bug, see #259.

This PR provides following:

• Separate bookkeeping of host and device memory updates for each vector in the multivector object.
• Add syncData overload to sync only one vector in multivector object. Closes Vector::getData function incorrectly syncs memory spaces. #259.
• Sets update flags correctly when vector elements are initialized using setToZero and setToConst methods. Closes Vector initialization functions do not set "updated" flags correctly #216.

Checklist:

• All tests pass. Code tested on
  • CPU backend
  • CUDA backend
  • HIP backend
• Code compiles cleanly with flags -Wall -Wpedantic -Wconversion -Wextra.
• The new code follows Re::Solve style guidelines.
• There are unit tests for the new code.
• The new code is documented.
• The feature branch is rebased with respect to the target branch.

[shakedregev]

I see not all checkmarks are there. Is this still a draft?

[kswirydo]

First -- this is not a bug. What you are adding is a new feature. There is nothing wrong with current approach except ... it is limiting in some cases.

[kswirydo]

I'm not sure if I follow the logic here. If the length of gpu_updated and 'cpu_updated' is k, shouldnt we just check every vector separately and copy, if needed?

[pelesh]

Good question! This function syncs the entire multivector at once. It can do that only if all vectors in the multivector have the same updated status. Otherwise, it should exit with an error. Next few lines check if it is the case. After that, it is safe to use all_gpu_updated.

The check itself may add some overhead. Perhaps it should be enabled only in debug mode?

[kswirydo]

Okay, I see what you are doing.

[shakedregev]

Looks good and tests pass, but maybe we want a test for this specifically? Separate PR?